Apple just announced a refresh of the Mac Studio with two new chips — the M5 Max and the M5 Ultra — and the latter is the company's most powerful chip ever. The M5 Ultra Mac Studio starts at $5,499, comes with up to 512GB of unified memory, and is the first desktop that lets you run frontier-size AI models entirely on your own machine at speeds that would make a cloud GPU cluster look slow.
It was announced alongside a new Mac mini with the 2nm M6 chip — but the M5 Ultra is where the real story is.
The M5 Ultra: a quad-die monster
The M5 Ultra uses Apple's next-generation UltraFusion interconnect to join two dual-die M5 Max chips into a quad-die architecture — something Apple has never done before. The result is a 36-core CPU (12 super cores, 24 performance cores), up to an 80-core GPU with Neural Accelerators in every core, and a 32-core Neural Engine.
Apple says it delivers up to 1.25x higher single-threaded performance and up to 1.3x higher multithreaded performance compared to the previous Ultra, along with up to 4.3x faster AI performance. The unified memory bandwidth hits 1.2TB/s — 50% higher than the M4 Ultra.
For context, the M4 Ultra topped out at 800GB/s of bandwidth. The M5 Ultra's 1.2TB/s is a generational leap that Apple has been building toward since it first connected two M1 Ultra dies together.
The M5 Max: pro-grade at a sane price
The M5 Max is a more accessible option. It features an 18-core CPU (6 super cores, 12 performance cores) and up to a 40-core GPU with Neural Accelerators. It supports up to 128GB of unified memory with 614GB/s of bandwidth.
Apple says the M5 Max delivers up to 3.9x faster AI performance and up to 1.8x faster graphics compared to the M4 Max. The base M5 Max Mac Studio starts at $2,499 with 36GB of memory and 512GB of storage — the same starting price as the M4 Max model after its $500 increase last June.
What matters for local AI
Here's why this is genuinely exciting for anyone who runs AI locally:
The M5 Ultra with 512GB of unified memory can load entire frontier models. A 70B parameter model in 4-bit quantization takes roughly 40GB of memory. A 405B model at 4-bit takes about 250GB. Even a 700B model at 4-bit (~440GB) would fit. This isn't theoretical — the M3 Ultra with 256GB already ran models like DeepSeek-R1 671B, and the M5 Ultra doubles the memory ceiling.
1.2TB/s of bandwidth means it's actually usable. Memory bandwidth is the bottleneck for LLM inference on Apple Silicon. At 1.2TB/s, the M5 Ultra can push a 70B model at easily 30-40 tokens per second, and a 405B model at maybe 5-8 tokens per second — not chat-speed for the 405B, but fast enough for serious experimentation, coding assistance, or batch processing. That's on your own machine. No API calls. No rate limits. No data leaving your building.
The M5 Max with 128GB is also a serious local AI machine. At 614GB/s, it runs 70B models at 15-20 tok/s and can comfortably fit a 70B model at 4-bit, or a 405B model at 2-bit (heavily quantized, but functional). For anyone who doesn't need the Ultra's 512GB, the $2,499 M5 Max is a compelling entry point.
The other upgrades
Both new Mac Studio models get a next-generation SSD architecture built on PCIe Gen 6 that Apple says is up to 2x faster than the previous generation. The I/O also gets Thunderbolt 5 (up to 120Gb/s) and Wi-Fi 7 / Bluetooth 6 via Apple's new N1 wireless chip for the first time.
Connectivity includes up to six Thunderbolt 5 ports, 10Gb Ethernet, HDMI 2.1, and support for up to eight displays.
Pricing and availability
- Mac Studio with M5 Max: starts at $2,499 (or $2,299 for education)
- Mac Studio with M5 Ultra: starts at $5,499 (or $5,099 for education)
Both are available to pre-order today and ship starting September 22.
Our take
Apple has spent the better part of four years building a local AI platform on Apple Silicon, and the M5 Ultra Mac Studio is its payoff. The $8,499 fully-configured M5 Ultra with 512GB of memory and 1.2TB/s of bandwidth is not a consumer device. It's a workstation-grade AI platform that makes a serious statement about what local AI can look like.
The question is whether Apple will lean into it. The company has been quietly promoting Macs for local AI development through MLX, but it hasn't marketed the Mac Studio as an AI workstation the way it markets the MacBook Air for students. If Apple were to position the M5 Ultra as the definitive local AI desktop — with clear benchmarks, developer tooling, and maybe even a "Mac Studio for AI" landing page — it would own a niche that no PC manufacturer can touch.
NVIDIA has the H100. AMD has the MI300X. But those cards start at $30,000+ and require data center infrastructure. The M5 Ultra Mac Studio does what they do — just not at the same scale — for $5,499, on a desk, using a fraction of the power.
For researchers, indie developers, and small labs that can't justify an H100 cluster but need to run frontier models, this is a game-changer. Apple may not be building the biggest AI engine in the world. But it might be building the most accessible one.